I am really enjoying the new app builder, and despite the loss of drag-and-drop editing capabilities, I have definitely experienced a net gain in productivity as the AI does a great job executing on my vision.
However, my ability to use the builder is being greatly diminished by its AI credit usage. Even basic tasks, like “Expand this section on the page” can consume 100 AI credits!
To build anything meaningful, I am blasting through the provided 3,000 credits/mo. That balance is typically gone in a day or two. And since the builder cannot forecast credit usage before you submit a prompt, you really just have to take a gamble whether or not you’ll be able to build your app before you run out of credits.
I am using the default model (Opus 4.8).
Is this unusual? Or just par for the course of AI app building in Retool? Would appreciate feedback or tips!
Using the new app builder with the default Opus 4.8 model, even basic tasks (e.g. 'expand this section') can burn ~100 AI credits, quickly exhausting the monthly allotment, and there's a question of whether this is expected or reducible.
AI Response
This is expected behavior rather than a bug: before any prompt runs, the builder pre-loads a large baseline context (resource schemas, source files, function outputs, dependency metadata, and conversation history), and credit consumption scales with reasoning complexity, output size, and total context processed. Several documented practices can meaningfully reduce burn: switch to a lighter model for simple edits (reserve Opus 4.8 for complex, multi-step changes), write precise prompts that name resources/components directly to avoid costly re-tries, and start a fresh thread when the task shifts so context doesn't accumulate. Credit and per-response usage can be tracked under Settings > Plans & Billing (Usage tab) and by hovering the icon below each agent response in the Chat tab.
Sources
Token management for app building | Retool Docs
Explains why baseline context is large and lists concrete ways to cut credit usage—model selection by task complexity, precise prompting, new threads, and usage monitoring. New models for building apps
Staff announcement confirming lighter-weight models (Claude Sonnet 5, the GPT-5.6 family, and Claude Fable 5) are now selectable, supporting the switch away from the default Opus 4.8 for simpler tasks.
The Community Team is testing out a new automation. Let us know if it's helpful (or not) by leaving a , , or . Or by marking this post as the "Solution"! Let us know if you have any feedback here.
I have the same exact issue. I can work for about a day and half with the 3k credits. Worse is that so many SDLC actions (deploying, debugging, version control, etc) are only accessible via prompt. So if you use all of your credits, then deploy, then see that the deployed app has a bug that didn’t show up in the editor - you can’t fix it unless you pay more or wait a month. Using your own API key obviously allows you to continue to work but in my experience it is even more expensive (see my post here New App Builder - First Impressions). The current MCP also does not help since it mostly just passes a prompt to the Retool agent.
My understanding is that the Retool team is working on some things that can be helpful for this although no timeline is communicated:
Allowing one to choose the “effort” level which currently uses provider defaults (for Anthropic this is “high”)
Enhancing the MCP so that a local agent (like a subscription) can be used to push code changes.
They are also hoping that model costs in general will come down which seems more like hope than a strategy.
Thank you for the response - glad to hear I’m not the only one bothered by this. Every time I prompt the builder to make a change, it starts by reading through the app to understand the context - then plans, then executes. But this becomes highly repetitive- re-scanning the app’s scope every time I ask for a change? Then, as the app grows, the problem compounds. Surely there must be a better way?
I’d like to try using GPT-5.6 Luna, but it’s only available on Enterprise (right)?
Hey there! Tanner from the product team here. Thanks for the thoughtful feedback on token/credit efficiency. We’re actively working on ways to make the product experience the right balance of quality & performance vs credit efficiency.
Luna is available for use on any paid plan– you only need to select it from the model selector dropdown!
Further, an admin can set a preferred model under Settings > AI so that all users in your org default to a preferred model.
Switching to ChatGPT models has tradeoffs - the builder does not communicate its steps like it does using Claude Sonnet. Higher error rate as well.
However, I’m not sure if this is due to Retool pushing some changes, but recent building has used far fewer credits than in the past even using Claude models. Glad to see it!
Yes, we have been working on this. Glad to see that you have felt a difference! We are also exploring dynamic model routing, including to OSS models (e.g. Qwen), which I think could drive this down further. Stay tuned!
, 
, or 
. Or by marking this post as the "Solution"! Let us know if you have any feedback 
